Current Trends in Text-to-Image Generation and AI Safety

The field of text-to-image (T2I) generation and AI safety is witnessing significant advancements aimed at addressing representativity biases and enhancing fairness. Recent developments focus on creating frameworks that evaluate and mitigate biases in T2I systems, ensuring that generated images are diverse, inclusive, and of high quality. These frameworks employ both human-based and model-based approaches to assess bias, with promising results indicating that model-based methods can effectively substitute human evaluations, potentially reducing costs and automating the process.

In parallel, there is a growing emphasis on reducing annotator bias in AI systems through innovative methods that elicit beliefs rather than direct judgments. This approach has shown potential in reducing systematic differences in annotations, thereby improving the generalizability of AI systems and preventing harm to underrepresented groups.

Additionally, the operationalization of AI safety principles is being reimagined through the vernacularization of harm taxonomies, which aims to make these frameworks more applicable to sector-specific contexts, particularly in underresourced or high-risk sectors. This participatory, decolonial practice is proving to be a valuable methodology for bridging gaps between formal taxonomies and local implementation needs.

The importance of diverse perspectives in evaluating the safety of generative AI is also being highlighted. Studies are demonstrating significant differences in how various demographic groups perceive and rate the safety of multimodal AI outputs, underscoring the need for inclusive safety evaluation processes.

Finally, advancements in prompt learning for fair T2I generation are addressing the quality degradation issues associated with current state-of-the-art methods. Novel techniques such as prompt queuing and attention amplification are being proposed to improve both the fairness and quality of generated images.

Noteworthy Papers

• Text-to-Image Representativity Fairness Evaluation Framework: Demonstrates effective bias capture in T2I systems and proposes model-based approaches as substitutes for human evaluations.
• Reducing annotator bias by belief elicitation: Proposes a method to reduce annotator bias through belief elicitation, showing consistent bias reduction in controlled experiments.
• Vernacularizing Taxonomies of Harm: Argues for the importance of vernacularizing harm taxonomies to enhance sector-specific AI safety operationalization.
• Insights on Disagreement Patterns in Multimodal Safety Perception: Highlights the need for diverse perspectives in AI safety evaluation through a large-scale study on multimodal safety perceptions.
• FairQueue: Rethinking Prompt Learning for Fair Text-to-Image Generation: Introduces techniques to improve both fairness and quality in T2I generation, outperforming current state-of-the-art methods.